Adversarial Code Review for ML Pipelines

Summary

Systematically hardens complex ML pipelines, such as BFCL, through iterative adversarial review rounds using external LLMs. This process tracks cumulative fixes and focuses on identifying subtle bugs that cause data corruption or evaluation integrity failures.